InclusionAI profile, products and governance

Worth watching if you follow open-weight models: Ant Group's in-house lab publishes MIT-licensed Ling, Ring and Ming releases from Hangzhou. Its newest, the Ling-3.1-flash preview announced 30 September 2026, is reachable only through hosted endpoints for now, while promised weights are still unpublished.

InclusionAI is Ant Group's open AGI research lab, founded in March 2025. It builds four model lines: Ling for efficient general use, Ring for agentic reasoning, Ming for multimodal generation and LLaDA for diffusion text. Most weights are MIT licensed. Its newest model, Ling-3.1-flash, is a 560B-parameter preview that had no published weights on 10 October 2026.

Founded: 2025 · HQ: Hangzhou, Zhejiang, China · Team: 50-200 · Funding: Internally funded as an Ant Group research division; no separate funding rounds, external investors, or valuation disclosed

About InclusionAI

InclusionAI was established in March 2025 as an open AGI research initiative inside Ant Group, the Hangzhou-based fintech company spun out of the Alibaba family that also owns Alipay. The lab is not separately incorporated; it operates as an internal Ant Group division and is led at the executive level by Zhengyu He, Ant Group's chief technology officer and a former academic researcher known for early work on GPU-accelerated maximum-flow algorithms. Richard Bian, who previously ran the AntOSS open-source program at Ant Group, leads product and growth for InclusionAI and its Ant Ling model line. The team is staffed with engineers and researchers who previously worked at OpenAI, Google and Meta's FAIR lab, alongside Ant Group's existing machine learning organization.

InclusionAI's output splits into four open-weight model lines. Ling is the token-efficient, non-reasoning foundation line: it started with Ling-1T, Ant's first trillion-parameter language model, unveiled in October 2025, and has since progressed through Ling-2.5-1T and Ling-2.6-1T, its current flagship. Ring is the explicit-reasoning counterpart, derived from the same Ling backbone and tuned for agentic workflows such as coding agents, tool use and long-horizon multi-step execution; Ring-2.6-1T is a trillion-parameter mixture-of-experts model with roughly 63 billion active parameters per token. Ming is InclusionAI's multimodal, any-to-any line built on top of Ling, culminating in Ming-Flash-Omni-2.0, which the lab describes as the first open model to unify speech, audio and music generation within one architecture. LLaDA is a separate diffusion-based language model line, most recently updated as LLaDA 2.1 with token-editing acceleration. Every model line is released under the MIT license, with weights distributed through Hugging Face, ModelScope and GitHub rather than sold as a hosted API.

The most recent release cycle has been dense. Ming-Flash-Omni-2.0 shipped in February 2026, the same month InclusionAI published Ling-2.5-1T and Ring-2.5-1T alongside real-world benchmark data, including an IMO 2025 gold-medal-standard score of 35 out of 42 for Ring-2.5-1T and a China Mathematical Olympiad score that beat the national team's qualifying cutoff. Ling-2.6-1T followed in April 2026 as the new head of the Ling line, and by September 2026 the lineup had grown further downmarket with Ling-3.0-tiny, a 7.9-billion-parameter model InclusionAI positions on the intelligence-per-active-parameter frontier, and Ling-3.0-flash-Fin, a finance-domain fine-tune of Ling-3.0-flash built with financial institutions and domain experts for investment research and spreadsheet-heavy workflows. The newest step is Ling 3.1 Flash, announced on 30 September 2026 as a preview: a mixture-of-experts model with about 560 billion total and 25 billion active parameters, aimed at coding agents and long tasks. It is reachable through hosted endpoints such as OpenRouter (262K context, listed free during the preview, accessed 10 October 2026), and InclusionAI has said it intends to open-source it soon, but no weights or licence were published as of 10 October 2026. Artificial Analysis scores it 41.1 on its Intelligence Index, which is the figure to weigh against the lab's own benchmark chart.

InclusionAI has not raised outside capital and has no disclosed valuation of its own; it is funded entirely out of Ant Group's internal budget as a strategic research program, the same way Google DeepMind or Meta's FAIR operate inside larger parent companies. No public reporting ties a specific dollar figure to InclusionAI's annual spend, distinguishing it from independently financed labs such as Anthropic or Mistral AI that raise and disclose named funding rounds.

The lab does not sell its models directly. Every major release ships as open weights under the MIT license, downloadable from Hugging Face and GitHub at no cost, which functions as a talent-recruitment and ecosystem-building strategy rather than a revenue line of its own. The commercial payoff for Ant Group is indirect: Ling and Ring-derived models feed Ant's own financial-services products, including the Ling-3.0-flash-Fin model built specifically for investment research and spreadsheet-heavy financial workflows, while the open releases build developer goodwill and recruiting pull against better-funded Western labs.

InclusionAI's Hugging Face organization page lists roughly 60 named researchers and engineers as of 2026, though that figure likely understates the lab's true headcount since it only reflects publicly credited contributors. Leadership sits with Zhengyu He at the executive level and Richard Bian on product and growth; beyond those two, InclusionAI does not publish an executive roster or board, consistent with its status as an internal division rather than a standalone company.

InclusionAI frames its mission as making compact, practical intelligence available to everyone, emphasizing efficient, deployable models over sheer scale for its own sake. Beyond language models, the lab has started publishing early work on Embodied AI, aiming to connect its digital models to physical-world robotics tasks, and treats every major release as a public technical report with benchmark data rather than a marketing-only announcement.

InclusionAI competes most directly with China's other open-weight labs. Against DeepSeek, its Ring reasoning line trades blows on agentic and math reasoning benchmarks while training at a reported 20% lower compute cost by using domestic Huawei Ascend and Alibaba-made chips instead of Nvidia hardware. Against Alibaba Cloud and its Qwen team (see what Qwen is for), Ling competes on general-purpose open mixture-of-experts models, and against Moonshot AI's Kimi line, InclusionAI's Ring models compete on long-context agentic reasoning. InclusionAI's edge is Ant Group's balance sheet and its access to Alipay-scale financial data, which none of those rivals can draw on directly; its gap is the absence of the independent governance and safety disclosure that Western frontier labs publish as standard.

Training on domestic Huawei Ascend and Alibaba-fabricated chips instead of export-controlled Nvidia GPUs is a deliberate response to US export controls on advanced AI chips to China, and Ant Group has said publicly that the substitution cut training costs by around 20% while matching Nvidia H800-class results. InclusionAI has not disclosed an EU AI Act compliance posture or a China-specific AI regulatory filing, and as a China-based, Ant Group-funded lab distributing MIT-licensed weights globally, it sits outside the compliance frameworks, such as SOC 2, HIPAA and GDPR data-processing agreements, that Western enterprise-facing AI vendors publish.

Who should not build on InclusionAI: teams that need a contracted vendor. The lab sells no hosted service of its own, publishes no SOC 2, ISO 27001 or GDPR paperwork, and has not disclosed an EU AI Act position, so a regulated buyer is better served by a provider with a trust center, such as Z.ai or Mistral AI on the hosted side. The open MIT weights are still the draw for self-hosting through Hugging Face, but Ling 3.1 Flash is not one of them yet. To compare the wider field, browse the models directory or answer a few questions in Smart Match for a shortlist matched to your use.

Mission

An earnest yet small wish to bring small and usable intelligence to the world.

Products

• Ling-3.1-flash (Preview MoE LLM for agents, hosted API only; open-source release promised, not yet published): https://openrouter.ai/inclusionai/ling-3.1-flash
• Ling-2.6-1T (Flagship open-weight non-reasoning MoE LLM): https://huggingface.co/inclusionAI/Ling-2.6-1T
• Ring-2.6-1T (Open-weight agentic reasoning MoE LLM): https://huggingface.co/inclusionAI/collections
• Ming-Flash-Omni-2.0 (Open-weight omni multimodal model): https://github.com/inclusionAI/Ming
• Ling-3.0-flash-Fin (Finance-domain fine-tuned LLM): https://huggingface.co/inclusionAI/Ling-3.0-flash-Fin
• LLaDA 2.1 (Open-weight diffusion language model): https://huggingface.co/inclusionAI/collections

Links

• Website
• GitHub
• Twitter
• Blog
• Docs